SWAMMA simulations use a limited-area version of the UK Met Office (UKMO) Unified Model (UM), based on the HadGEM3-RA regional climate model previously tested at various resolutions over Africa (Moufouma-Okia and Jones, 2015). The UM is designed to function across a wide range of spatial and temporal scales and is used for meteorology and climate research as well as operational numerical weather prediction. The UM (version 8.2 is used here) consists of a dynamical core (Davies et al., 2005; Staniforth et al., 2006) which describes evolution of the atmosphere as a non-hydrostatic, fully compressible fluid. Model levels are terrain which is close to the surface but relaxes to smooth, parallel levels at height. The model has a fixed Eulerian grid but utilises the semi-implicit, semi-Lagrangian time stepping to advect variables (allowing for mass conservation). Physics packages include a two-stream radiation code (Edwards et al., 2012), the Joint UK Land Environment Simulator (JULES) land surface exchange scheme (Best, 2005; Best et al., 2011), boundary layer turbulence (Lock and Edwards, 2012), cloud microphysics (Wilkinson, 2012) and convection (Stratton et al., 2009). The SWAMMA simulations use a limited area set-up with a domain encompassing all of West Africa (approximately 0–35∘ N and 23∘ W–35∘ E). Simulations are conducted at horizontal grid spacings of 4, 12 and 40 km, all having 70 levels in the vertical. The 12 and 40 km models have a rigid model lid at a height of 80 km, while the 4 km version has a rigid lid at 40 km height. The differences in vertical spacing between the 4 km simulations and the rest of the simulations means that the height of the lowest model level is also different (approximately 2.5 m for the 4 km simulations and 10 m for the rest). Model levels are concentrated in the lower atmosphere to better represent meteorological processes. The simulations are initialised at the beginning of the simulation period on 1 May and run until the end of the simulation period without being reinitialised. As such the interior of the model is able to behave in a similar “free-running” way to regional climate simulations, and the model monsoon system is able to develop without the strict constraints of analysed conditions. This allows for the characteristics of the modelled monsoon to arise and highlight model errors that are likely present in other similarly constrained simulations. The lateral boundary conditions (horizontal winds and potential temperature) are updated every hour and produced by performing global simulations using the UM on an N216 (∼ 60 km) grid (also version 8.2). Global simulations are initialised every 6 h using European Centre for Medium-Range Weather Forecasts (ECMWF) operational analysis data. Sea surface temperatures for the limited-area SWAMMA runs are updated every 6 h and obtained by regridding ECMWF operational analysis data (as above). Land surface features are handled by the JULES land surface exchange scheme with some features being described by invariant ancillary files (vegetation fraction) and others evolving due to simulated conditions (soil moisture). The UM configuration used for SWAMMA is not dissimilar to that used by the Cascade simulations (a series of simulations over West Africa for a 40-day period in summer 2006 using different grid spacing nested into one another and both parameterised and explicit convection; for further details see Birch et al., 2014; Pearson et al., 2014), and settings in many of the model physics sections, notably the representation of convection, were adopted from there. The SWAMMA simulations are longer than those in Cascade, running from initialisation from 1 May to 30 September 2011 (153 days): this spans an entire monsoon season, allowing for investigation of the development of the West African Monsoon (WAM) in an unprecedented way.

Another important improvement on the Cascade simulations is the inclusion of prognostic interactive dust in the SWAMMA simulations. This allows for the investigation of the dust-raising and transportation characteristics of the model under varying resolutions and convection options, as well as assessing the radiative impact that dust has on the WAM system. The dust scheme used is that within the Coupled Large-scale Aerosol Simulator for Studies in Climate (CLASSIC; Johnson et al., 2011) scheme, in which dust particles are assumed to be spherical and transported in the atmosphere as six independent tracers undergoing dry deposition through turbulent mixing and gravitational settling as well as wet deposition through washout from precipitation. Dust emissions are calculated during each model time step using prognostic model fields. The dust emission scheme utilises the widely used algorithm of Marticorena and Bergametti (1995) to calculate horizontal flux in each of nine bins with boundaries at 0.0316, 0.1, 0.316, 1.0, 3.16, 10.0, 31.6, 100., 316 and 1000 µm radius (the largest three of the size modes are only active in saltation processes). Each of the six dust size bins is treated independently by the radiation scheme with spectral properties being calculated from Mie theory. The horizontal dust flux for dust particles in each size bin is calculated as a function of the cube of the surface friction velocity (U*), the bare-soil fraction in the grid box (shown in Fig. 1), the mass fraction of soil particles available at the surface, and a threshold surface friction velocity (Ut*) below which dust is not mobilised. Assuming that the lowest model level (for wind speed) is with the turbulent boundary near the surface, then the calculation of U* should be insensitive to different vertical grid spacings. This seems to be the case in the SWAMMA simulations with there being no clear relationship between different grid spacings across the suite of simulations and the model diagnostic U* values. The threshold value (Ut*) is a function of soil moisture in the top layer (10 cm thick in the model) and the clay fraction in the grid box, such that emissions are inhibited for wet soils (further details in Woodward 2011; Ackerley et al., 2012). Dust emission models may be tuned by adjusting coefficients by which U* and the top-level soil moisture are multiplied with a global tuning factor. Here the values used are 1.6, 0.5 and 2.5 respectively, and values were not adjusted for different model grid spacings in order to make a fair comparison between the model run at different resolutions. The Harmonized World Soil Database (FAO, 2012) is used to determine soil texture and thus the fractions of clay, silt and sand available in each surface grid box for the dust emission scheme. Dust fields are initialised from zero and drop to zero on the lateral boundaries (so that no dust enters the domain at the boundaries). Surface infra-red emissivity is changed from the JULES default value over bare soil (0.97) to 0.9 for these experiments, as this is more realistic over the Sahara (Ogawa and Schmugge, 2004). As described above there is no explicit use of preferential dust sources; however, where soil characteristics and surface roughness are favourable the threshold friction velocity over bare soil (Ut*) can be reduced, allowing for favourable emission conditions in particular regions.

Within the framework described above, eight simulations are conducted which comprise the SWAMMA model suite. The main variable factors between the simulations are grid spacing, representation of convection and radiatively interactive mineral dust (see Table 1). In simulations with parameterised convection the convective scheme in the UM is switched on (Stratton et al., 2009). This scheme is based on a convective available potential energy (CAPE) closure method, where high CAPE values are identified and tendencies are determined to reduce this over a given timescale. In the simulations with explicit convection the convective parameterisation has effectively been switched off by increasing the CAPE closure timescale to a point at which CAPE depletion by the parameterisation is insignificant. These models employ a Smagorinsky-style subgrid-scale mixing in all three dimensions (3DS in Table 1) with mixing length constants chosen as those found optimal for the 12 and 4 km models in Cascade (0.05 and 0.1 respectively). In the simulations with radiatively active dust, mineral dust emitted from the surface within the simulations influences the radiation budget via its direct radiative effect (scattering and absorbing solar and thermal radiation); cloud microphysical effects are not included. While dust is present in the radiatively inactive simulations it does not influence the radiation budget or the evolution of the model meteorology. Comparing simulations with different convection types but without dust effects (e.g. 12P and 12E in Table 1) highlights the impact of resolved convection on dust generation without complications of feedbacks through dust–radiation interactions. We focus on the latter in this paper, although here we note that effects of interactive dust on both dust uplift itself and thermodynamics are far smaller than those that change the convection (not shown).

Wind speed observations from several in situ observation platforms are compared with simulated wind speeds. Data from five stations are used, these are Fennec automatic weather stations (AWSs) 134 (23.5∘ N, 3.0∘ W) and 138 (27.4∘ N, 3.0∘ W), the Fennec flux tower deployed at Bordj Badji Mokhtar (BBM; 28.3∘ N, 0.9∘ E) and African Monsoon Multidisciplinary Analysis (AMMA) “Couplage de l'Atmosphère Tropicale et du Cycle Hydrologique” (CATCH) AWSs at Agoufou (15.3∘ N, 1.5∘ W) and Kobou (14.7∘ N, 1.5∘ W). Different dust uplift mechanisms occur at different times of the day. The advantage of these observations compared to routine synoptic observations is their high temporal resolution (which allows for resolution of the diurnal cycle) as well as the geographical spread of stations across the Sahel and Sahara, which are generally very poorly observed. To compare between simulations and observations taken at different heights, all winds are adjusted to 2 m height using the wind profile power law u=ur(z/zr)α, where ur is wind speed reference height (zr), z is the height to be adjusted, and α is a stability coefficient (nominally 0.143; Touma, 1977; Roberts et al., 2017).

To investigate the nature of mesoscale convective systems seen in observations and those generated in convection-permitting simulations a storm-tracking approach has been adopted. The algorithm used is based on that of Stein et al. (2014) and has been modified for use on both simulations and observations (Crook et al., 2018). The algorithm can be applied to either rainfall or brightness temperatures to track convective systems over West Africa. Storm clusters are identified through the use of a threshold and by grouping contiguous cells. These are then followed in time using a fractional overlap method (0.6 overlap threshold) to track storm cells, allowing for both cell splitting and merging. If a storm has no overlapping cells from the previous time step, then it is a new initiation. When a storm has no overlapping cells in the next time step, it is a dissipation. For splits the cell with the greatest overlap retains its storm ID, while other cells are said to have split and are given new storm IDs (parent IDs are recorded). Similarly, for merging, the cell from the previous time step with the greatest overlap with the resultant cluster maintains its ID and any other cells with smaller overlaps are said to have merged and take the ID of the cell with the largest overlap. For this study it was decided that a brightness temperature approach, using a threshold of -40 ∘C would be best suited. This is due to the use of hourly data for tracking, where clouds give a greater overlap and therefore a chance of tracking between time steps. This is also because we are interested in systems for which rain does and does not reach the surface, since both situations can produce haboobs, making rainfall tracking less reliable. Therefore, for this study, hourly brightness temperatures calculated from both simulated and observed (SEVIRI channel 9, 10.8 µm) outgoing long-wave radiation has been used to track systems, giving information about storm triggering, locations, size and storm lifetime.

Comparisons of dust AODs with observations are frequently used to verify (and in many cases, tune) dust models (Huneeus et al., 2011, 2016). This is because AOD observations from satellites are now available at high temporal and spatial resolutions, unlike observations of dust emissions and concentrations. However, within a modelling framework AOD is very much an end product, requiring not only accurate representations of all the physical processes involved in dust emission, transport and deposition to achieve realistic dust loadings but also accurate representations of particle size distribution and spectral optical properties. For example, in the SWAMMA experiments, although extinction per unit mass is greatest for particle size division 2 (0.1–0.3 µm mean radius), dust mass is maximum in division 4 (1–3 µm mean radius), and total extinction for dust is dominated by particles in size division 3 (0.3–1 µm mean radius). On the other hand, models with very similar AODs can have very different dust emissions due to compensating differences in deposition, transport or particle size distribution (Kinne et al., 2003; Ocko and Ginoux 2016; Evan et al., 2014).

The dust loadings in the SWAMMA experiments (5–6 Tg May to September seasonal mean for the whole domain) are at the low end of, but not outside, the range reported by other modelling studies (this of course could be resolved by tuning total emissions, but would not affect the systematic model biases we investigate here); Huneeus et al. (2011) reviewed 15 global models within the AeroCom project and found global loadings ranged between 7 and 30 Tg, of which ∼ 70 % has been estimated to be attributable to the Sahara (Luo et al., 2003). All versions of the model here are initialised with zero dust and found to be spun-up within 5–10 days; for ease of analysis and consistency with presentation, monthly means for May are presented here for the whole month with no special treatment of the spin-up period (it should be noted that even including spin-up May results in dust and AOD values in excess of any other simulated month).

Figure 2 displays the monthly mean (May–September) AODs at 550 nm from the MODIS Terra satellite with the dust AOD from all the SWAMMA models excluding dust radiative effects (4E, 12E, 12P and 40P from Table 1). Here the model AOD at 10:00 UTC has been selected to provide a better time match for the Terra data which overpass the region at approximately 10:30 LST. As Fig. 2 shows monthly mean values of AOD at approximately 10:00 UTC for both simulations and satellite retrievals, we believe that this is a good comparison with which the overall differences in the spatial distribution of dust in both reality and the simulations can be judged. It is clear that the models are very similar across all resolutions and all feature a maximum over the Bodélé depression (∼ 18∘ N, 19∘ E) in all months, in common with the MODIS data. However, apart from in May, the models all have insufficient dust over the central Sahara and a strong maximum over the west coast at ∼ 20∘ N, which is not evident in the MODIS data. These are common features of many models (e.g. Fig. 2 of Todd and Cavazos-Guerra, 2016, Fig. 6 of Ridley et al., 2012) and, as pointed out by Evan et al. (2014) in a multi-model CMIP5 comparison study, may have many contributory factors, including a poor representation of soil texture, moisture and vegetation cover, and deficiencies in model surface winds. The focus of this study is to see if any improvement can be achieved by resolving convection explicitly, since haboobs are known to be a key uplift mechanism in the summertime central Sahara. Figure 3 therefore compares the spatial correlations and biases (at model grid points) for the 12 km simulations with explicit and parameterised convection (12E and 12P) relative to the MODIS AODs, broken down into specific regions as shown by the boxes in Fig. 1b: northern Sahara (NS, 25–30∘ N), the Sahara (SA, 15–25∘ N), the Sahel (SL, 10–15∘ N) and the Guinea coast (GC, 5–10∘ N). As mentioned in Sect. 2.2.1 MODIS AOD retrievals are compared with the nearest corresponding times from simulations to reduce errors from diurnal variations. Overall, where they are significant (>0.5), correlations are positive, except for the Guinea coast region in June where dust loads are lower and, as noted in Sect. 2.2.1, MODIS data are anomalous. Correlations are high (>0.5) for the northern Sahara throughout the season and also in May in the Sahel, and lower at other locations and times, which are when moist convection and haboobs are known to be most active. Differences between the explicit and parameterised versions of the model are small, with the parameterised version generally having slightly better correlations with MODIS except for July–September in the Sahara. Despite the low correlations in the summertime Sahara, this is the region where we would look to find improvements in dust in the convection-permitting simulations. This is due to the expectation that in this region there are areas with surface characteristics that allow for the deflation of dust, as well as the additional uplift process that is expected to be represented (haboobs). The model AOD biases relative to MODIS data are predominantly negative and have the greatest magnitude to the south of the SWAMMA region, consistent with the model producing too little dust there (although the maximum bias of ∼ -0.55 in June in the Guinea coastal region is where the MODIS data are anomalous). Exceptions to this are the northern Sahara and Sahara in May, where biases are positive but small, suggesting too little dust in regions and seasons in which moist convection is most active but too much prior to the monsoon onset and close to the Atlantic coast. Where there are differences between explicit and parameterised simulations, the explicit version mostly has larger biases than the parameterised model, although the differences are small.

All the SWAMMA models lack the AOD maximum evident in the MODIS data from June to August in the central Sahara. We therefore examine factors affecting the dust emission to see why this might be. Figure 4 shows the monthly mean (May–September) dust AODs (for all hours), with the corresponding dust emissions, surface friction velocity over bare soil (U*) and soil moisture in the top 10 cm soil layer for the 12 km explicit convection model (12E). Reference to the clay fractions in Fig. 1 is also helpful, as it is a factor in the vertical dust flux equation. Areas with high clay fraction (up to a maximum value of 0.2) have the potential to produce the most dust in dry conditions. However it is also the case that high clay soils are more sensitive to soil moisture, with higher soil moisture values impeding emission. South of approximately 15∘ N, emission of dust is negligible due to the very small bare-soil fractions (see Fig. 1) and higher soil moisture values (Fig. 4); although the JULES surface exchange scheme includes a seasonal climatology of fractional leaf area index (LAI), the fraction of each land type, including bare soil, is fixed. It is known that there is a strong seasonal cycle in vegetation over the Sahel (Mougin et al., 2009) with summertime dust emission from haboobs during the early monsoon season (June–August; Klose et al., 2010; Knippertz and Todd, 2012), and even cold pools from congestus clouds can lead to visible dust uplift in June (Marsham et al., 2009). This fixed bare-soil fraction therefore means that seasonal dust emission from the Sahel cannot be realistically represented in this configuration of the UM. Figure 4 also shows that, in the model, emissions are strong over the Bodélé, the west coast and central Algeria, where the highest U* values coincide with regions of high clay fraction and low soil moisture. To see the impact of the choice of convection scheme, Fig. 5 shows the differences between the 12 km explicit and parameterised models for the same variables. There is a clear mid-season switch in the AOD, emissions and U* 12E-12P differences such that they are generally (over the whole SWAMMA area) much more negative in July–September than in May–June. Differences in soil moisture have relatively little impact on the dust emissions, because they are mainly to the south of the region, where the bare-soil fraction is small. However, it should be noted that the soil moisture available to be input to the emission scheme is the top 10 cm mean soil moisture; this is likely to have a buffering effect on emissions, as in reality the skin soil moisture controls dust emission. This has a much faster timescale for drying and reaching a level appropriate for mineral dust deflation: Gillette et al. (2001) reports sediment from a dry lake in California being raised 10–30 min after rainfall, and Bergametti et al. (2016) reports Sahelian surfaces taking less than 12 h to fully recover their dry-sand transport potential. Some areas of increased U* and dust emissions are evident for the 12E model in the central Sahara (∼ 20∘ N, 0∘ E) in June, and to a lesser extent in July, but these do not produce any overall decrease in the MODIS bias of the model for the SA region in Fig. 3 due to compensating increases elsewhere.

Figure 6 summarises the seasonal trends of AOD and in factors affecting the dust AOD in the 12E and 12P models for the northern Sahara (NS), Sahara (SA) and Sahel (SL) regions. We see that for both simulations the AODs (monthly mean values of all available times) are poorly simulated with their highest values in May, whereas MODIS AOD increases from May to a maximum in July (for NS and SA) and June (for SL). The trend in AOD shown by MODIS retrievals is consistent with the summertime northwards advance of the monsoon, rainfall and haboobs (Marsham et al., 2008). MODIS AOD data from 2006–2008 in Fig. 2 of Ridley et al., 2012 indicate that this pattern is robust and not unique to 2011, indicating that simulations are missing a key dust-generating mechanism providing a maximum in June–July. Additionally we see that the explicit convection version generally performs worse than the parameterised version in this respect. Analysing the contributory factors, the trend in model AOD follows the trend in dust load, as expected (note that loads plotted are regional totals scaled by a factor of 5 for NS and SA, and 10 for SL). The dust loads generally follow the trend in dust emissions, except for May–June in the Sahel where the dust load is boosted by advection from the Sahara. Dust emission trends are strongly driven by the friction velocity (U*) in NS and SA throughout the season, where soil moisture values are too low to have much influence (except for SA in August where the monsoon rains encroach on the region) and any trend in modelled soil moisture cannot explain the decrease in modelled dust from May to September. For the Sahel the pattern is different, but with lower U* and much higher soil moisture values combining to drastically reduce emissions and dust loads as the monsoon season evolves. In the Sahara and Sahel friction velocity values are generally lower, and soil moisture values higher in the convection-permitting than the parameterised model, leading to lower dust emissions, loadings and AODs (since the monsoon is further north in the explicit run, not shown but consistent with Marsham et al., 2013 and Birch et al., 2014). For the northern Sahara the explicit version has lower soil moisture and higher U* than the parameterised in May–June, leading to higher dust AODs which exceed the MODIS values; however, this is not sustained over the rest of the season and AOD biases are negative for July–September.

The explicit treatment of convection is known (from Cascade; Marsham et al., 2011 and Heinold et al., 2013) to have a strong impact on the representation of haboobs in the UM, but here it does not impact the dust fields significantly. We therefore continue our investigation with an evaluation of the near-surface winds (a strong controlling factor in the emission of dust) in both simulations and observations, to further explain why explicitly permitting haboobs has such a small impact on the modelled dust AODs.

The hypothesis that explicit convection would produce significant differences in the dust field for the SWAMMA simulations has been shown to be incorrect. One potential cause of this is the possibility that the simulated surface winds do not change very much from one simulation to another. Figure 7 shows the distribution of wind speeds adjusted to an observation height of 2 m using the wind profile power law (Touma, 1977; Roberts et al., 2017) at a number of locations in the Sahel and Sahara for all four simulations (4E, 12E, 12P and 40P) as well as observed winds. There is close agreement in the maximum frequency of occurrence in the simulations at each of the stations, with the observations up to 3 m s-1 lower. The largest differences occur in the Sahel. The advantage of showing the distributions on a logarithmic y axis is that the frequency of rare high wind speed events can be examined. While the frequency of such events might be low, the non-linear nature of wind speed to dust uplift (above a threshold) means that they dominate dust uplift (Cowie et al., 2015). The maximum simulated winds (when accounting for height adjustment) are similar to values shown for 10 m in ALADIN simulations, with parameterised convection in Chaboureau et al. (2016). While the ALADIN simulations with explicit convection show an increased frequency in very strong winds (due to convective cold pools), the SWAMMA simulations show little change in the frequency of very strong winds. It is plausible that models from other centres are also likely to respond differently to the UM. F-134 and BBM in the southern Sahara have both observed high wind speed events that are significantly underrepresented in all of the simulations. This is particularly important as both these stations are located in a region where haboobs are known to be significant and in the seasonal maxima of AOD that can be seen in Fig. 2 (top row) but is absent in all SWAMMA simulations (and CMIP5 simulations). In 3 of the 5 stations the maximum wind values produced in the 4 km convection-permitting simulations are lower than that seen in the 12 km convection-permitting simulation. While this has not been investigated any further here, it highlights a possible scale dependence for the maximum strength of winds that are generated by convective cold pools in convection-permitting models and is possibly linked to scaling of updraught and downdraught column cross sections, although Huang et al. (2018) suggests that 4 km subgrid spacing in the cold pool itself is not critical for dust-generating winds.

In order to investigate haboob winds, Fig. 8a, b show the anomaly of 10 m wind speed cubed composited around column maximum rainfall rates greater than 1 mm h-1 for the 12 km simulations with parameterised and explicit convection (over the region 15∘ W to 20∘ E, 10–30∘ N, where it is expected that if a cold pool were to occur it could feasibly raise dust). The anomaly is calculated as the difference from average wind speed cubed values calculated for each simulated month and time of day to reduce the effects of the seasonal and diurnal cycles. The period for the composite average covers the time at which the rain threshold is met and the following 6 h. This highlights the production of convective cold pools in the convection-permitting simulation which have wind speed cubed values in excess of the average for that time of day and season. The peak in the centre of the composite domain in Fig. 8b and the absence of a peak in Fig. 8a indicate that cold pools are indeed present in the 12 km explicit convection simulation and that these features are absent in the parameterised version. Another important feature of the cold-pool anomaly shown in Fig. 8b is the extent of the positive anomaly field. Although this cannot provide a direct measurement of the size of cold pools in the simulation it clearly indicates that they can reach very large sizes (in excess of 300 km radii). With cold pools of this size it might be expected that there would be a noticeable impact on the uplift of dust and therefore the distribution of AOD. Figure 8c, d are the same type of composite as Fig. 8a, b but for the maximum 10 m wind speed recorded within the rainfall plus the 6 h window described above. This gives greater information about the actual strength of the winds generated by the presence of convectively generated cold pools. The maximum wind speed composite for the parameterised convection simulation (Fig. 8c) indicates generally weaker winds than seen in the explicit simulation. However, the maximum cold-pool winds (which clearly show a positive anomaly in Fig. 8b) are relatively weak, reaching maximum composite values of between 6 and 8 m s-1. This is lower than the mean maximum wind seen in Provod et al. (2016) for observed Sahelian cold pools of 8 to 10 m s-1. As the values in the centre (and just north of centre) of the domain would be most likely to be affected by almost all cold pools generated, it would be expected that there would be only a minor effect of reducing the wind speed values via compositing. With this in mind, it is surprising that the maximum value measured would be approximately 8 m s-1 as this represents only a minor exceedance of (or even a failure to exceed) the approximate 7–8 m s-1 dust uplift threshold used in many emission schemes (Marticorena et al., 1997). This suggests that, although there are clearly cold pools being generated in the convection permitting simulation, and these cold pools produce anomalously strong winds, they are not as strong as might be expected. Certainly the strongest cold-pools winds, which are known to generate the extreme winds in the southern Sahara and Sahel stations, are missing in the models (Fig. 7).

The unchanging overall frequency of different wind speeds (Fig. 7) and the presence of convectively generated cold pools (Fig. 8) can be combined with the findings of Marsham et al. (2013) that up to 50 % of dust emission in the summertime central Saharan hotspot occurs at night due to haboobs. This highlights the need to compare diurnal cycles of the different simulations. Figure 9 shows the diurnal cycle of dust uplift potential (DUP; Marsham et al., 2011) for all simulated months for the five sites for which there are observations. The northern Sahara station, F-138, has a similar development of the diurnal cycle across the 5 simulated months: in both observations and simulations the highest DUP values tend to occur during the day with much lower values at night, and in some months there is a maximum at 09:00 UTC, likely from the breakdown of the nocturnal LLJ. This is as expected given that F-138 is too far north to be strongly or regularly influenced by the cold pools spreading deep into the Sahara. The low night-time values reflect the development of a stable nocturnal boundary layer, which breaks down due to surface heating during daylight hours. F-134 and BBM in the Saharan box show a clearer peak from LLJ breakdown at approximately 09:00 UTC. At F-134 (in both observations and simulations) this process is the dominant feature throughout the entire season. However, further south at BBM, the observations suggest that the morning peak in DUP is similar in magnitude, with an evening peak, in agreement with Marsham et al. (2013). This second peak in DUP associated with haboobs is not well represented in the simulations with the 12 km explicit and 4 km explicit simulations having different diurnal cycles with regard to the evening peak. This is possibly caused by the simulations failing to produce cold pools of sufficient strength as far north as BBM. However, the evening peak at BBM cannot be wholly attributed to cold pools. This is due to the fact that there is a similar, yet smaller, peak present in the simulations, with parameterised convection in June. This is feasibly the impact of the daily night-time surge of the monsoon flow, which is stronger in the parameterised simulations than the explicit simulations (consistent with Birch et al., 2014).

At the Sahelian stations of Agoufou and Kobou, the diurnal cycle in May (Fig. 9p, u) is similar to that seen in the Sahara with a morning LLJ peak in DUP and largely similar diurnal behaviour across all simulations. However, by June there is evidence of divergent behaviour between the simulations. At Agoufou and Kobou (Fig. 9q, v) there is an evening peak in DUP at 16:00–21:00 UTC, which is more pronounced in convection-permitting simulations. This evening peak grows more pronounced at these stations from July to August. This evening peak is also particularly noisy: this behaviour is what would be expected from high DUP values associated with cold pools due to their production of very high wind values that last on timescales <1 h (for observations at a fixed point). When combined, these features mean that an average diurnal cycle produced over a relatively short period of time (1 month) will not produce a smooth evening peak in DUP. Concomitantly the convection-permitting simulations also have a reduction in size of the morning NLLJ DUP peak (consistent with Marsham et al., 2011). As shown earlier in Fig. 7 there is little change in the overall distribution of modelled wind speed with explicit convection, showing how the increased evening winds are compensated for by the decreasing morning winds overall.

Given that it has been shown that convective cold pools are present and are likely to be responsible for a significant modification of the diurnal cycle of winds in the Sahel and as far north as BBM it follows that there should be some modification in the uplift and transport of dust. Figure 10 shows the monthly mean diurnal cycle in dust emissions from the 12 km simulations for the five stations. Although there is some evidence of an evening increase in emissions in the convection-permitting model at BBM in June–August, consistent with the DUP in Fig. 9, this is insufficient to significantly change or improve the dust load and AOD. Dust emissions at stations in the Sahel (Agoufou and Kobou) are reduced in the explicit version: this is likely to be due to the increased soil moisture in that region (as demonstrated in Fig. 6). In addition to such limits imposed by the surface characteristics on the uplift of dust in the model, it is also possible that there is some behaviour of convective storms and their associated cold pools that means that they do not lift dust; for example the wrong size, lifetime or location. This is examined in the next section.

To investigate the nature of the storms that are responsible for the generation of cold pools, a storm-tracking approach has been used. This takes advantage of the availability of satellite observations of outgoing long-wave radiation from which the brightness temperature can be easily derived, and tracking is performed on features with a brightness temperature below -40 ∘C. This means that direct comparisons of the storms produced in the convection-permitting simulations can be made with those identified through satellite retrievals. To reduce the number of events that were considered and to highlight the impact of larger events which dominate observed dust uplift in the central summertime Sahara (Marsham et al., 2013; Allen et al., 2013), only storms that reached a threshold value in size (approximately 5000 km2) were considered. These systems will be referred to as mesoscale convective systems (MCSs) hereafter. The total number of MCSs between the 4 and 12 km convection-permitting simulations and the observations is not dissimilar, having 14,082, 15,555 and 12,843 respectively; however, the storm track densities (number of times a storm track is centred over a specific region on a 0.25∘ grid) in Fig. 11a–c shows that there is a greater density of events in both of the simulations compared to the observations. This is likely due to the generally enhanced lifetime of MCSs seen in simulations (Fig. 11e). The spatial distribution of MCS track density (based on storm centres) indicates that their latitudinal position in simulations compared to reality is approximately correct, with MCSs (and therefore cold pools) commonly occurring as far as 17∘ N, indicating that the positioning of the MCSs is not the driving factor behind the lack of dust raised here. However, the higher frequency of very large storms in SEVIRI imagery compared to explicit simulations (Fig. 11d), the generally weak cold-pool winds identified in Fig. 8d and the missing tail of high winds in Fig. 7b, c, suggest that the region affected by large cold pools has an underrepresentation of cold-pool winds in convection-permitting simulations.

Figure 11e shows the distributions of the MCS duration (to the nearest hour), highlighting the fact that the 4 and 12 km simulations have MCSs that last longer on average than those in SEVIRI; it is only storms that live beyond 30 h for the 4 km simulations and 47 h for the 12 km simulations that the frequency of occurrence first drops below the values seen from observations. This abundance of events (even MCSs that are smaller than those observed) and the fact that convective cold pools are clearly being produced in the simulations (despite their reduced strength) suggests that the lack of emission in the simulation south of 17∘ N cannot, however, be entirely attributed to smaller MCSs producing smaller and weaker cold pools.

In interpreting the storm-track-based analysis discussed above, it is useful to examine sample images of observed and modelled large storms. Figure 12 is a case study of a large cold-pool event that occurred on 23 August 2011. It was well represented in the 4 km simulation in that the timing and location of initiation of the system was roughly correct, after which a large MCS developed and produced a cold pool which spread north and west into the Sahara. Although we do not necessarily expect an accurate one-to-one correspondence between observed and modelled storms this far into the simulation, the case shown does share key similarities and is one of the larger modelled storms from the simulated period. The cold pool in the simulation can be seen through both the elevated friction velocity over bare soil as well as the spreading of air away from the MCS shown in the 10 m wind vectors. Similarly, the cold pool generated in reality can be identified through the occurrence of arc clouds along the leading edge of the cold pool and the magenta colour that identifies raised dust within the cold pool in the SEVIRI RGB false colour dust images. The impact that this cold pool has on dust is assessed through the daytime averages of the dust AOD from the 4 km simulation and the SEVIRI AERUS-GEO AOD product. There is clearly a strong AOD signal associated with the cold pools in both measures. However, the signal in the simulation is dwarfed by the high levels over the western part of the domain (at least partially associated with erroneously high uplift over the Western Sahara). In the SEVIRI AERUS-GEO product the AOD feature in the central Sahara is comparable in magnitude to the transported plume over the Atlantic and is much more clearly linked to uplift caused by strong near-surface winds associated with the passage of a convective cold pool. This is consistent with the maximum mean hourly observed wind on this day at BBM being 11.4 m s-1 and the maximum instantaneous modelled cold-pool wind being 7.5 m s-1.